A symmetric prior knowledge based deep learning model for intracerebral hemorrhage lesion segmentation

Area of Science:

• Medical Imaging and Artificial Intelligence
• Neurology and Neurosurgery
• Radiology

Background:

• Accurate localization and classification of intracerebral hemorrhage (ICH) lesions are critical for patient treatment and prognosis.
• Computed tomography (CT) is a key imaging modality for diagnosing ICH.
• Deep learning offers potential for automated lesion segmentation and diagnosis.

Purpose of the Study:

• To develop and evaluate a novel symmetric prior knowledge-based deep learning model, Sym-TransNet, for segmenting ICH lesions in CT images.
• To assess the diagnostic accuracy, sensitivity, and specificity of Sym-TransNet for ICH detection.

Main Methods:

• A symmetric Transformer network (Sym-TransNet) was designed for ICH lesion segmentation.
• A dataset of 1,157 ICH patients and 200 healthy subjects was used for training, validation, and testing.
• The DICE coefficient was employed to evaluate segmentation performance against ground truth; comparisons were made with existing deep learning methods.

Main Results:

• Sym-TransNet achieved a DICE coefficient of 0.716 for ICH lesion segmentation in a test set of 200 CT images.
• Specific subtype segmentation showed varying DICE coefficients: IPH (0.784), IVH (0.680), EDH (0.359), SDH (0.534), and SAH (0.337).
• Overall diagnostic accuracy, sensitivity, and specificity for ICH were 91.25%, 98.50%, and 84.00%, respectively.

Conclusions:

• Sym-TransNet demonstrates superior performance in segmenting and identifying ICH lesions compared to recent deep learning methods.
• The model provides stable and efficient ICH diagnosis, indicating significant clinical application potential.
• Ablation studies confirmed the effectiveness of each component within Sym-TransNet for improving segmentation.